CipherGuard-V4.2-SAPO-0.6B-GGUF

This is the GGUF (F16) conversion of CipherGuard-V4.2-SAPO-0.6B for use with llama.cpp and compatible inference engines.

CipherGuard is a proprietary Small Language Model (SLM) developed internally by SmartM2M for real-time safety classification and sensitive information detection in LLM-powered applications.

With only 596M parameters, CipherGuard is designed for edge deployment and low-latency inference — serving as a lightweight guardrail layer that can be integrated into any LLM pipeline.

Available Files

File	Quant	Size	Description
CipherGuard-V4.2-SAPO-0.6B-f16.gguf	F16	~1.2 GB	Full precision — no quality loss

Classification Standard (V4.2)

This model follows the "actual data required" classification principle:

A message is Unsafe only when it contains, discloses, or requests actual sensitive data (real numbers, real keys, real records). Merely mentioning a sensitive keyword or asking about a process does not make it Unsafe.

Text	Label	Reason
주민등록번호 (keyword only)	Safe	Keyword mention without actual data
주민등록번호는 950101-1234567 입니다	Unsafe	Contains actual SSN value
계좌번호 입력란이 어디에 있나요?	Safe	Procedural question, no actual data
계좌번호 110-123-456789 로 송금해주세요	Unsafe	Contains actual account number

Key Features

• Bilingual: Native support for Korean and English
• Real-time classification: Safe / Unsafe / Controversial with category breakdown
• Edge-ready: 596M params, runs efficiently on CPU or single GPU
• High precision: 92.6% Unsafe precision — minimal false positives
• GGUF format: Direct use with llama.cpp, Ollama, LM Studio, and more

Usage

With llama.cpp

llama-cli -m CipherGuard-V4.2-SAPO-0.6B-f16.gguf \
  -p "<|im_start|>user\n제 주민등록번호는 900101-1234567 입니다<|im_end|>\n<|im_start|>assistant\n" \
  -n 64 --temp 0.2

With llama-cpp-python

from llama_cpp import Llama

llm = Llama(model_path="CipherGuard-V4.2-SAPO-0.6B-f16.gguf")
output = llm.create_chat_completion(
    messages=[{"role": "user", "content": "제 주민등록번호는 900101-1234567 입니다"}],
    max_tokens=64,
    temperature=0.2,
)
print(output["choices"][0]["message"]["content"])
# Safety: Unsafe
# Categories: PII

With Ollama

# Create a Modelfile
echo 'FROM ./CipherGuard-V4.2-SAPO-0.6B-f16.gguf' > Modelfile
ollama create cipherguard -f Modelfile
ollama run cipherguard "계좌번호 110-123-456789 로 송금해주세요"

Output Format

Safety: <Safe|Unsafe|Controversial>
Categories: <comma-separated list or None>

Safety Labels

Label	Description
Safe	No actual sensitive data present (keyword mentions alone are Safe)
Unsafe	Contains actual sensitive data values (real numbers, keys, records, etc.)
Controversial	Ambiguous content requiring human review

Supported Categories

PII · Financial Information · Health/Biometric Information · Account Information · API Key/Access Token · Internal Document · HR/Payroll Information · Credential/Secret · Location/Device Information · Organization/Project Information · Legal Information · Real Estate Information

Evaluation

Evaluated on 2,138 samples labeled with V4.2 criteria (keyword only = Safe, keyword + actual value = Unsafe).

Metric	Value
Safety Accuracy (3-class)	90.08%
Binary Accuracy	90.69%
PII Category Accuracy	91.4%

Per-Class Performance

Class	Precision	Recall	F1
Safe	89.5%	91.3%	90.4%
Unsafe	92.6%	92.0%	92.3%
Controversial	4.2%	2.6%	3.2%

Model Details

Attribute	Value
Base Model	CipherGuard-V4.2-SAPO-0.6B
Architecture	CipherGuard SLM (596M params)
Training Method	SFT (DoRA) + SAPO
GGUF Quantization	F16 (full precision)
Languages	Korean, English
License	Apache 2.0

Limitations

• Optimized for Korean and English — other languages may have reduced accuracy
• Category detection beyond PII/API Key/Internal Document needs improvement
• Controversial class detection is weak (F1 3.2%)
• Designed as a classification layer, not a general-purpose language model

Citation

@misc{cipherguard2026,
  title={CipherGuard: Lightweight Safety Classification SLM for Enterprise Guardrails},
  author={SmartM2M Bandung},
  year={2026},
  url={https://huggingface.co/smartm2m-bdg/CipherGuard-V4.2-SAPO-0.6B}
}

---

Developed by SmartM2M Bandung — AI Safety & Guardrails Division